multicollinearity Multicollinearity
What is. Let H = the set of all the X (independent) variables. Let Gk = the set of all the X variables except Xk. The formula for standard errors is then
2 s 1 RYH *y 2 (1 RX k Gk ) * ( N K 1) s X k
sbk =
=
2 s 1 RYH *y Tolk * ( N K 1) s X k
= Vif k *
2 s 1 RYH *y ( N K 1) s X k
The bigger R2XkGk is (i.e. the more highly correlated Xk is with the other IVs in the model), the bigger the standard error will be. Indeed, if Xk is perfectly correlated with the other IVs, the standard error will equal infinity. This is referred to as the problem of multicollinearity. The problem is that, as the Xs become more highly correlated, it becomes more and more difficult to determine which X is actually producing the effect on Y. Also, recall that 1 - R2XkGk is referred to as the Tolerance of Xk. A tolerance close to 1 means there is little multicollinearity, whereas a value close to 0 suggests that multicollinearity may be a threat. The reciprocal of the tolerance is known as the Variance Inflation Factor (VIF). The VIF shows us how much the variance of the coefficient estimate is being inflated by multicollinearity. The square root of the VIF tells you how much larger the standard error is, compared with what it would be if that variable were uncorrelated with the other X variables in the equation. For example, if the VIF for a variable were 9, its standard error would be three times as large as it would be if its VIF was 1. In such a case, the coefficient would have to be 3 times as large to be statistically significant. Causes of multicollinearity Improper use of dummy variables (e.g. failure to exclude one category) Including a variable that is computed from other variables in the equation (e.g. family income = husbands income + wifes income, and the regression includes all 3 income measures) In effect, including the same or almost the same variable twice (height in feet and height in inches; or, more commonly, two different operationalizations of the same identical concept) The above all imply some sort of error on the researchers part. But, it may just be that variables really and truly are highly correlated.
MulticollinearityPage 1
Consequences of multicollinearity Even extreme multicollinearity (so long as it is not perfect) does not violate OLS assumptions. OLS estimates are still unbiased and BLUE (Best Linear Unbiased Estimators) Nevertheless, the greater the multicollinearity, the greater the standard errors. When high multicollinearity is present, confidence intervals for coefficients tend to be very wide and tstatistics tend to be very small. Coefficients will have to be larger in order to be statistically significant, i.e. it will be harder to reject the null when multicollinearity is present. Note, however, that large standard errors can be caused by things besides multicollinearity. When two IVs are highly and positively correlated, their slope coefficient estimators will tend to be highly and negatively correlated. When, for example, b1 is greater than 1, b2 will tend to be less than 2. Further, a different sample will likely produce the opposite result. In other words, if you overestimate the effect of one parameter, you will tend to underestimate the effect of the other. Hence, coefficient estimates tend to be very shaky from one sample to the next.
Detecting high multicollinearity. Multicollinearity is a matter of degree. There is no irrefutable test that it is or is not a problem. But, there are several warning signals: None of the t-ratios for the individual coefficients is statistically significant, yet the overall F statistic is. If there are several variables in the model, though, and not all are highly correlated with the other variables, this alone may not be enough. You could get a mix of significant and insignificant results, disguising the fact that some coefficients are insignificant because of multicollinearity. Check to see how stable coefficients are when different samples are used. For example, you might randomly divide your sample in two. If coefficients differ dramatically, multicollinearity may be a problem. Or, try a slightly different specification of a model using the same data. See if seemingly innocuous changes (adding a variable, dropping a variable, using a different operationalization of a variable) produce big shifts. In particular, as variables are added, look for changes in the signs of effects (e.g. switches from positive to negative) that seem theoretically questionable. Such changes may make sense if you believe suppressor effects are present, but otherwise they may indicate multicollinearity. Examine the bivariate correlations between the IVs, and look for big values, e.g. .80 and above. However, the problem with this is One IV may be a linear combination of several IVs, and yet not be highly correlated with any one of them Hard to decide on a cutoff point. The smaller the sample, the lower the cutoff point should probably be.
MulticollinearityPage 2
Ergo, examining the tolerances or VIFs is probably superior to examining the bivariate correlations. Indeed, you may want to actually regress each X on all of the other Xs, to help you pinpoint where the problem is. A commonly given rule of thumb is that VIFs of 10 or higher (or equivalently, tolerances of .10 or less) may be reason for concern. This is, however, just a rule of thumb; Allison says he gets concerned when the VIF is over 2.5 and the tolerance is under .40. In SPSS, you get the tolerances and vifs by adding either the VIF or COLLIN parameter ro the regression command; in Stata you can use the vif command after running a regression, or you can use the collin command (written by Philip Ender at UCLA). Look at the correlations of the estimated coefficients (not the variables). High correlations between pairs of coefficients indicate possible collinearity problems. In SPSS, you can get this via the BCOV parameter; in Stata you get it by running the vce, corr command after a regression. Sometimes eigenvalues, condition indices and the condition number will be referred to when examining multicollinearity. While all have their uses, I will focus on the condition number. The condition number () is the condition index with the largest value; it equals the square root of the largest eigenvalue (max) divided by the smallest eigenvalue (min), i.e.
=
max min
When there is no collinearity at all, the eigenvalues, condition indices and condition number will all equal one. As collinearity increases, eigenvalues will be both greater and smaller than 1 (eigenvalues close to zero indicate a multicollinearity problem), and the condition indices and the condition number will increase. An informal rule of thumb is that if the condition number is 15, multicollinearity is a concern; if it is greater than 30 multicollinearity is a very serious concern. (But again, these are just informal rules of thumb.) In SPSS, you get these values by adding the COLLIN parameter to the Regression command; in Stata you can use collin. CAUTION: There are different ways of computing eigenvalues, and they lead to different results. One common approach is to center the IVs first, i.e. subtract the mean from each variable. (Equivalent approaches analyze the standardized variables or the correlation matrix.) In other instances, the variables are left uncentered. SPSS takes the uncentered approach, whereas Statas collin can do it both ways. If you center the variables yourself, then both approaches will yield identical results. If your variables have ratio-level measurement (i.e. have a true zero point) then not centering may make sense; if they dont have ratio-level measurement, then I think it makes more sense to center. In any event, be aware that authors handle this in different ways, and there is sometimes controversy over which approach is most appropriate. I have to admit that I dont fully understand all these issues myself; and I have not seen the condition number and related statistics widely used in Sociology, although they might enjoy wider use in other fields. See Belsley, Kuh and Welschs Regression Diagnostics: Identifying Influential Data and Sources of Collinearity (1980) for an in-depth discussion.
MulticollinearityPage 3
Dealing with multicollinearity Make sure you havent made any flagrant errors, e.g. improper use of computed or dummy variables. Increase the sample size. This will usually decrease standard errors, and make it less likely that results are some sort of sampling fluke. Use information from prior research. Suppose previous studies have shown that 1 = 2*2. Then, create a new variable, X3 = 2X1 + X2. Then, regress Y on X3 instead of on X1 and X2. b3 is then your estimate of 2 and 2b3 is your estimate of 1. Use factor analysis or some other means to create a scale from the Xs. In fact, you should do this anyway if you feel the Xs are simply different operationalizations of the same concept (e.g. several measures might tap the same personality trait). In SPSS you might use the FACTOR or RELIABILITY commands; in Stata relevant commands include factor and alpha. Use joint hypothesis testsinstead of doing t-tests for individual coefficients, do an F test for a group of coefficients (i.e. an incremental F test). So, if X1, X2, and X3 are highly correlated, do an F test of the hypothesis that 1 = 2 = 3 = 0. It is sometimes suggested that you drop the offending variable. If you originally added the variable just to see what happens, dropping may be a fine idea. But, if the variable really belongs in the model, this can lead to specification error, which can be even worse than multicollinearity. It may be that the best thing to do is simply to realize that multicollinearity is present, and be aware of its consequences.
SPSS Example. Consider the following hypothetical example:
MATRIX DATA VARIABLES = Rowtype_ Y X1 X2/ FORMAT = FREE full /FILE = INLINE / N = 100. BEGIN DATA. MEAN 12.00 STDDEV 3.00 CORR 1.00 CORR .24 CORR 0.25 END DATA. REGRESSION 10.00 10.00 5.00 5.00 .24 .25 1.00 0.95 0.95 1.00
matrix = in(*) /VARIABLES Y X1 X2 /DESCRIPTIVES /STATISTICS DEF TOL BCOV COLLIN TOL /DEPENDENT Y /method ENTER X1 X2 .
MulticollinearityPage 4
Regression
Descriptive Statistics Y X1 X2 Mean 12.000000 10.000000 10.000000 Std. Deviation 3.0000000 5.0000000 5.0000000 N 100 100 100
Correlations Pearson Correlation Y X1 X2 Y 1.000 .240 .250 X1 .240 1.000 .950 X2 .250 .950 1.000
b Variables Entered/Removed
Model 1
Variables Entered a X2, X1
Variables Removed .
Method Enter
a. All requested variables entered. b. Dependent Variable: Y Model Summary Model 1 R R Square .250a .063 Adjusted R Square .043 Std. Error of the Estimate 2.9344301
a. Predictors: (Constant), X2, X1
ANOVAb Model 1 Sum of Squares 55.745 835.255 891.000 df 2 97 99 Mean Square 27.872 8.611 F 3.237 Sig. .044a
Regression Residual Total
a. Predictors: (Constant), X2, X1 b. Dependent Variable: Y
Coefficientsa Unstandardized Coefficients B Std. Error 10.492 .666 .015 .189 .135 .189 Standardized Coefficients .026 Beta .226 Collinearity Statistics Tolerance VIF .098 .098 10.256 10.256
Model 1
(Constant) X1 X2
t 15.765 .081 .717
Sig. .000 .935 .475
a. Dependent Variable: Y
MulticollinearityPage 5
a Coefficient Correlations
Model 1
Correlations Covariances
X2 X1 X2 X1
X2 1.000 -.950 .036 -.034
X1 -.950 1.000 -.034 .036
a. Dependent Variable: Y
a Collinearity Diagnostics
Model 1
Dimension 1 2 3
Eigenvalue 2.855 .136 .010
Condition Index 1.000 4.589 16.964
Variance Proportions (Constant) X1 X2 .02 .00 .00 .98 .02 .02 .00 .98 .98
a. Dependent Variable: Y
Note that X1 and X2 are very highly correlated (r12 = .95). Of course, the tolerances for these variables are therefore also very low and the VIFs exceed our rule of thumb of 10. The t-statistics for the coefficients are not significant. Yet, the overall F is significant. Even though both IVs have the same standard deviations and almost identical correlations with Y, their estimated effects are radically different. The correlation between the coefficients for X1 and X2 is very high, -.95 The condition number (SPSS does not explicitly report it, but it is the largest of the condition indices) is 16.964. This falls within our rule of thumb range for concern. Again, this is based on the uncentered variables; if I thought centering was more appropriate I would just need to change the means of X1 and X2 to 0. (Doing so produces a condition number of 6.245, as Stata confirms below.) The sample size is fairly small (N = 100).
All of these are warning signs of multicollinearity. A change of as little as one or two cases could completely reverse the estimates of the effects.
MulticollinearityPage 6
Stata example. It is easy to do the same analysis as above using Stata. We use the corr, regress, vif, vce, and collin commands.
. use http://www.nd.edu/~rwilliam/stats2/statafiles/multicoll.dta, clear . corr y x1 x2, means (obs=100) Variable | Mean Std. Dev. Min Max -------------+---------------------------------------------------y| 12 3 4.899272 18.91652 x1 | 10 5 -1.098596 23.10749 x2 | 10 5 -.0284863 23.72392 | y x1 x2 -------------+--------------------------y| 1.0000 x1 | 0.2400 1.0000 x2 | 0.2500 0.9500 1.0000 . reg y x1 x2, beta Number of obs F( 2, 97) Prob > F R-squared Adj R-squared Root MSE = = = = = = 100 3.24 0.0436 0.0626 0.0432 2.9344
Source | SS df MS -------------+-----------------------------Model | 55.7446181 2 27.872309 Residual | 835.255433 97 8.61088075 -------------+-----------------------------Total | 891.000051 99 9.00000051
-----------------------------------------------------------------------------y| Coef. Std. Err. t P>|t| Beta -------------+---------------------------------------------------------------x1 | .0153846 .1889008 0.08 0.935 .025641 x2 | .1353847 .1889008 0.72 0.475 .2256411 _cons | 10.49231 .6655404 15.77 0.000 . -----------------------------------------------------------------------------. vif Variable | VIF 1/VIF -------------+---------------------x1 | 10.26 0.097500 x2 | 10.26 0.097500 -------------+---------------------Mean VIF | 10.26 . vce, corr | x1 x2 _cons -------------+--------------------------x1 | 1.0000 x2 | -0.9500 1.0000 _cons | -0.1419 -0.1419 1.0000
MulticollinearityPage 7
. * Use collin with uncentered data, the default. (Same as SPSS) . collin x1 x2 if !missing(y) Collinearity Diagnostics SQRT RVariable VIF VIF Tolerance Squared ---------------------------------------------------x1 10.26 3.20 0.0975 0.9025 x2 10.26 3.20 0.0975 0.9025 ---------------------------------------------------Mean VIF 10.26 Cond Eigenval Index --------------------------------1 2.8546 1.0000 2 0.1355 4.5894 3 0.0099 16.9635 --------------------------------Condition Number 16.9635 Eigenvalues & Cond Index computed from scaled raw sscp (w/ intercept) Det(correlation matrix) 0.0975 . * Use collin with centered data using the corr option . collin x1 x2 if !missing(y), corr Collinearity Diagnostics SQRT RVariable VIF VIF Tolerance Squared ---------------------------------------------------x1 10.26 3.20 0.0975 0.9025 x2 10.26 3.20 0.0975 0.9025 ---------------------------------------------------Mean VIF 10.26 Cond Eigenval Index --------------------------------1 1.9500 1.0000 2 0.0500 6.2450 --------------------------------Condition Number 6.2450 Eigenvalues & Cond Index computed from deviation sscp (no intercept) Det(correlation matrix) 0.0975
collin is a user-written command; type findit collin to locate it and install it on your machine. Note that, with the collin command, you only give the names of the X variables, not the Y. If Y has missing data, you have to make sure that the same cases are analyzed by the collin command that were analyzed by the regress command. There are various ways of doing this. By adding the optional if !missing(y) I told Stata to only analyze those cases that were NOT missing on Y. By default, collin computed the condition number using the raw data (same as SPSS); adding the corr parameter makes it compute the condition number using centered data. [NOTE: coldiag2 is yet another Stata routine that can give you even more information concerning eigenvalues, condition indices, etc.; type findit coldiag2 to locate and install it.]
MulticollinearityPage 8
Incidentally, assuming X1 and X2 are measured the same way (e.g. years, dollars, whatever) a possible solution we might consider is to simply add X1 and X2 together. This would make even more sense if we felt X1 and X2 were alternative measures of the same thing. Adding them could be legitimate if (despite the large differences in their estimated effects) their two effects did not significantly differ from each other. In Stata, we can easily test this.
. test x1 = x2 ( 1) x1 - x2 = 0 F( 1, 97) = Prob > F = 0.10 0.7484
Given that the effects do not significantly differ, we can do the following:
. gen x1plusx2 = x1 + x2 . reg y x1plusx2 Source | SS df MS -------------+-----------------------------Model | 54.8536183 1 54.8536183 Residual | 836.146432 98 8.53210645 -------------+-----------------------------Total | 891.000051 99 9.00000051 Number of obs F( 1, 98) Prob > F R-squared Adj R-squared Root MSE = = = = = = 100 6.43 0.0128 0.0616 0.0520 2.921
-----------------------------------------------------------------------------y| Coef. Std. Err. t P>|t| [95% Conf. Interval] -------------+---------------------------------------------------------------x1plusx2 | .0753846 .0297309 2.54 0.013 .0163846 .1343846 _cons | 10.49231 .6624892 15.84 0.000 9.17762 11.807 ------------------------------------------------------------------------------
The multicollinearity problem is obviously gone (since we only have one IV). As well see later, we could do the same thing in SPSS, but it would take a few additional steps. As noted before, factor analysis could be used for more complicated scale construction. Special Case: Models with nonlinear or nonadditive (interaction) terms. Sometimes models include variables that are nonlinear or nonadditive (interactive) functions of other variables in the model. For example, the IVS might include both X and X2. Or, the model might include X1, X2, and X1*X2. We discuss the rationale for such models later in the semester. In such cases, the original variables and the variables computed from them can be highly correlated. Also, multicollinearity between nonlinear and nonadditive terms could make it difficult to determine whether there is multicollinearity involving other variables. Consider the following simple example, where X takes on the values 1 through five and is correlated with X2.
MulticollinearityPage 9
* Example of multicollinearity with polynomial terms. DATA LIST FREE / X XSquare. BEGIN DATA. 1 1 2 4 3 9 4 16 5 25 END DATA. CORRELATIONS /VARIABLES=x xsquare /PRINT=TWOTAIL NOSIG /MISSING=PAIRWISE .
Correlations
Correlations X X Pearson Correlation Sig. (2-tailed) N Pearson Correlation Sig. (2-tailed) N 1 . 5 .981** .003 5 XSQUARE .981** .003 5 1 . 5
XSQUARE
**. Correlation is significant at the 0.01 level (2-tailed).
As we see, the correlation between X and X2 is very high. High correlations can likewise be found with interaction terms. It is sometimes suggested that, with such models, the original IVs should be centered before computing other variables from them. You center a variable by subtracting the mean from every case. The mean of the centered variable is then zero (the standard deviation, of course, stays the same). The correlations between the IVs will then often be far smaller. For example, if we center X in the above problem by subtracting the mean of 3 from each case before squaring, we get
* Now we center the variable first. DATA LIST FREE / W WSquare. BEGIN DATA. -2 4 -1 1 0 0 1 1 2 4 END DATA. CORRELATIONS /VARIABLES=W Wsquare /PRINT=TWOTAIL NOSIG /MISSING=PAIRWISE .
MulticollinearityPage 10
Correlations
Correlations W W Pearson Correlation Sig. (2-tailed) N Pearson Correlation Sig. (2-tailed) N 1 . 5 .000 1.000 5 WSQUARE .000 1.000 5 1 . 5
WSQUARE
As you see, the extremely high correlation we had before drops to zero when the variable is centered before computing the squared term. Well discuss nonlinear and nonadditive models later in the semester. Well also see other reasons why centering can be advantageous. In particular, centering can make it a lot easier to understand and interpret effects under certain conditions. The specific topic of centering is briefly discussed on pages 30-31 of Jaccard et als Interaction Effects in Multiple Regression. Also see pp. 35-36 of Aiken and Wests Multiple Regression: Testing and Interpreting Interactions. Note, incidentally, that centering is only recommended for the IVs; you generally do not need or want to center the DV. Multicollinearity in Non-OLS techniques The examples above use OLS regression. As we will see, OLS regression is not an appropriate statistical technique for many sorts of problems. For example, if the dependent variable is a dichotomy (e.g. lived or died) logistic regression or probit models are generally better. However, as Menard notes in Applied Logistic Regression Analysis, much of the diagnostic information for multicollinearity (e.g. VIFs) can be obtained by calculating an OLS regression model using the same dependent and independent variables you are using in your logistic regression model. Because the concern is with the relationship among the independent variables, the functional form of the model for the dependent variable is irrelevant to the estimation of collinearity. (Menard 2002, p. 76). In other words, you could run an OLS regression, and ignore most of the results but still use the information that pertained to multicollinearity. Even more simply, in Stata, the collin command can generally be used regardless of whether the ultimate analysis will be done with OLS regression, logistic regression, or whatever. In short, multicollinearity is not a problem that is unique to OLS regression, and the various diagnostic procedures and remedies described here are not limited to OLS.
MulticollinearityPage 11